About me

AI/ML Engineer with 5+ years of hands‑on experience shipping scalable systems in NLP, LLMs, and agentic automation. I’ve built production‑grade RAG chatbots, applicant scoring tools, and intelligent retrieval platforms used by 100+ daily users. I also contribute to ML research in Vision Transformers, Graph Neural Networks, Reinforcement Learning, and Robotics.

5+ yrs experience LLMs & Agents RAG & Vector Search FastAPI & Docker

My Technical Expertise

  • ml icon

    Machine Learning Engineering

    Deploying scalable ML/LLM systems in production: RAG, recommendation engines, ATS scoring, NLP pipelines, and real-time inference with observability.

  • automation icon

    AI Automation

    Agentic automation with multi-agent flows, structured validation, and MCP-backed retrieval that trims latency and cost while improving reliability.

  • knowledge icon

    Knowledge Graphs

    Modeling relationships for intelligent retrieval and reasoning via Neo4j/GraphQL integrations and curated graph insights.

  • recommendation icon

    Recommendation Systems

    Personalization at scale: vector search, rule-based re-ranking, dynamic multi-filter matching, and session-aware re-scoring.

  • data icon

    Data Engineering

    Robust pipelines for crawling, processing, and vectorizing large datasets, supported by Redis caching and observability.

  • quant icon

    Quantitative & Financial Modeling

    Algorithmic trading, risk modeling, and RL research with legality-aware action masking and custom Gym environments.

Resume

Work Experience

  1. Machine Learning Engineer — CanApply, Montreal, Canada

    Mar 2023 — Present
    • Architected a production-grade RAG chatbot with modular multi-agent flows, user profiling, multilingual handling, structured validation, and MCP-backed retrieval via Qdrant with web search fallback.
    • Integrated token‑efficient knowledge augmentation cutting inference time by 2×; added Redis session persistence, streaming tokens, and Sentry.io observability for scalable LLM conversations.
    • Engineered an AI program recommendation system: semantic vector search + dynamic rule-based multi-filter matching via Qdrant with FastAPI endpoints for real‑time personalization.
    • Implemented session persistence, efficient pagination, dynamic re-filtering on user updates, and full production monitoring, achieving sub‑800ms semantic matching under real-time load.
    • Led an internal ATS engine with agentic evaluation across 100 academic‑fit rubrics, coordinated by a central reasoning agent for feedback.
    • Developed a Scrapy-powered academic crawler indexing 500k+ pages from 54 institutions into Elasticsearch (keyword/partial) and Qdrant (semantic) for RAG agents.
    • CanSpider — full‑stack faculty discovery/classification pipeline: LLM‑assisted plan generation → Scrapy + Playwright crawl → Kafka events → LLM classification & digestion → MySQL upserts → admin UI for review/matching.
    • Implemented cost‑aware LLM orchestration: lightweight classification gates heavier extraction; cached completions in Qdrant to cut cost and improve reproducibility.
    • Shipped resilient streaming with aiokafka worker pool, backpressure, large‑payload producer, and graceful shutdown; secured APIs with JWT/API keys and idempotent POSTs.
    • Built a Next.js + Tailwind operational dashboard to launch/pause crawls, monitor logs/stats, restart consumers, trigger indexing, and bulk upsert digests.
    • Deployed Dockerized AI microservices on Hetzner VPS; optimized FastAPI performance and Redis caching for sub-second responses at 100+ sustained API req/min; added Sentry.io for real-time error tracking.
  2. Data Engineer and Software Developer at Sepanta IT Co., Mashhad, Iran

    Feb 2018 — Nov 2022
    • Built and optimized data pipelines for TSE stocks and cryptocurrencies, achieving sub-500ms latency for real-time data retrieval with MongoDB.
    • Developed a high-performing NLP model for Persian news sentiment analysis, utilizing Hazm toolkit and TF-IDF vectorization, with a macro average F1 score of 0.85.
    • Designed and deployed a backend API for a Telegram-based market analysis platform, using FastAPI and MongoDB for real-time market insights and user profiling.
    • Created algorithmic trading strategies, including option bonds pricing using Black-Scholes and Momentum Trading, and clustering models for industry performance analysis.
    • Programmed a real-time cryptocurrency triangular arbitrage system on Binance, handling over 1000 trading pairs with under 100ms latency, deployed on AWS EC2.

Research & Publications

  1. ViTHL: Vision Transformer-Based Hybrid Localization for Humanoid Robots — RoboCup Symposium 2025

    Contributed ViT + UKF/MCL integration for real-time robot localization (48ms @30Hz, 79% fewer divergence).

Education

  1. Honours Bachelor of Science in Computer Science

    Lakehead University, Thunder Bay, Canada

Certifications

  1. Deep Learning Specialization, Coursera

    2022
  2. Artificial Intelligence, IPM Advanced School on Computing

    2021
  3. MikroTik Certified Network Associate, Forat Technical Institute

    2020
  4. C# Programming, Forat Technical Institute

    2020
  5. How to Write a Good Research Proposal, Mashhad University of Medical Sciences

    2019
  6. CompTIA Network+

    2013

Honors & Awards

  1. Ahwazi Young Investigator Award – Behavioral Neuroscience

    2016

    Cognitive Sciences and Technologies Council.

  2. 3rd Place – National Cognitive Neuroscience Competition (IPM)

    2016

    Stimulus-response latency analysis in cockroach nervous system.

  3. 4th Place – Sharif University Robotics Competition (Smart Gardeners League)

    2012

    Designed and implemented autonomous gardening robotics strategies.

Tech Stack and Skills

  • Python
  • R
  • C/C++
  • C#
  • SQL
  • Bash
  • PyTorch
  • TensorFlow
  • Scikit-learn
  • Keras
  • Stable-Baselines3
  • WandB / MLflow
  • Hugging Face
  • OpenAI API
  • SpaCy
  • LangChain
  • Model Context Protocol
  • Pandas
  • NumPy
  • Spark
  • Airflow
  • Selenium
  • Scrapy
  • RDKit
  • Docker
  • Redis
  • FastAPI / Flask
  • GraphQL
  • Kubernetes
  • Sentry
  • Git / CI/CD
  • AWS S3 / GCP
  • MariaDB
  • Qdrant
  • Neo4j
  • PostgreSQL
  • MongoDB
  • Elasticsearch

Selected Projects

  1. Trellion — AI Hiring Platform with Leaderboard

    2025

    Two-sided AI platform to help candidates and recruiters hire faster. Candidates create a profile, auto-build/optimize resumes per job, generate cover letters, apply in seconds, and complete a personalized AI interview. Recruiters post jobs and review a ranked leaderboard generated from interview + profile signals to shortlist top finalists without scanning hundreds of resumes. Distinctive mechanism: per‑job AI interview, multi‑category scoring (skills/behavior/communication), candidate report card, and a live, transparent leaderboard (with Copilot on job pages and visible match%).

    LLM AgentsRecommendationGraphQLPostgreSQLFastAPI
  2. AI-Powered Pet-Caregiver Matching System — PAWSOME Concierge

    2025

    Production-grade FastAPI backend to match pet-care requests with qualified caregivers. Combined rule-based filtering with a structured 22‑question rubric scored by an LLM agent. Integrated Redis, MySQL, OpenAI + Google APIs, and Sentry.io. Powers 100+ monthly bookings.

    FastAPIRedisMySQLLLM Agents
  3. Petra Animal Review Sentiment & Topic Analysis

    2025

    Dual-pipeline NLP analyzing 4,200+ tourist snippets on animal welfare using rule-based and transformer models (85% agreement, r=0.58). UMAP + HDBSCAN topic clustering across 20 labeled themes.

    TransformersUMAPHDBSCAN
  4. RL-Driven Blackjack Simulator with Action-Masking

    2025

    Gym-compatible blackjack environment with partial observability and legality-aware actions. Trained PPO & DQN agents (DQN loss rate 49.4%, outperforming house edge) using Stable-Baselines3.

    RLStable-Baselines3Gym
  5. GNN-Based Pharmacological Interaction Engine

    2024

    Predicted drug interaction severity using SMILES graphs and RDKit descriptors. Modeled 16k+ structures with GAT, GIN, and MPNN variants.

    GNNRDKitPyTorch